Research Of Generative Object Tracking Method Under Deep Learning Framework

As one of the research hotspots in information process and computer vision,object tracking technique plays an important role in various issues,i.e.,security surveillance,intelligent transportation,visual navigation as well as military reconnaissance.Taking the initial location state of an interesting object as reference,tracking aims to accurately predict the location state of the object in each subsequent frame.Despite great progress has been realized in the past few years,there are still several difficult problems.It is a very challenging topic to achieve high-quality tracking in complex real scenarios,due to background clutter,motion blurring,illumination variation,occlusion,et al.As a result,exploiting both accurate and reliable tracking methods has great theoretical significance and application value.With the rapid development of artificial intelligence,deep learning models have been gradually applied in object tracking domain,and make a breakthrough in the field.The tracking methods under deep learning framework has achieved pretty promising tracking performance,especially the generative approaches based on Siamese networks.Therefore,this paper first carefully analyzes the basic theories of deep learning and object tracking,and then focuses on studying the generative object methods under deep learning architecture.The main research contents and innovative achievements are as follows:1.Existed Siamese trackers generally cannot ensure precision and robustness simultaneously,and lack of adaptivity to the appearance variations of object.To address these problems,this paper presents a hierarchical-aware attentional Siamese network for visual tracking.In offline training phase,both the position-aware and the appearance-aware training schemes are adopted.The former introduces massive spatial transformations while training the shallow networks,which is very useful to learning positive-aware features.The latter utilizes abundant appearance variations to optimize deep network layers,which is helpful to capture the appearance-aware features.By combining these training schemes,the robustness and precision of tracker would be increased obviously.Besides,an effective feature selection module is designed based on channel and spatial attention mechanisms,which is capable of helping tracker to adapt to the appearance changes more robustly.As last,the experimental results demonstrate that the proposed method is pretty state-of-the-art and effective,whcih can achieve both accurate and robust tracking performance.2.Traditional fusion method has no ability to combine multi-layer convolutional features adaptively,so they are hard to greatly improve the tracking capacity of Siamese networks.To address the problem,this paper proposes a novel feature fusion module for Siamese networks by studying residual learning theories.Specifically,the network employs the high-level features from deep layers as direct input to recognize the object globally with the abstract semantic information,and explores the shallow-layer features through residual channel to refine the object state with the local detailed information.To avoid the degradation problem,this paper also presents an ensemble training scheme for our tracker,in which different loss functions are introduced to individually optimize the Siamese and the fusion modules.The experimental results demonstrate that the proposed residual fusion network has an ability to greatly improve the tracking capacity of Siamese networks.3.Due to lack of efficient online update mechanism,typical Siamese networks are difficult to stably and reliably track an object in complicated scenes.To tackle the issue,this paper first proposes a novel two-stage one-shot learner by analyzing the one-shot learning scheme in Siamese networks,which can combine diverse-stage object samples to predict the network weights of the primary classifier.Then,an updatable Siamese tracker is designed based the learner.Especially,an extra input branch is used to capture the object appearance features in subsequent stages.Next,we construct a feature update module to combine the diverse template features,which can generate a fusing template with more patterns.The fusing module would help tracker to identify the tracked object in the search region more robustly.The experimental results prove the effectiveness of the proposed method,which can finish both stable and robust tracking in complex scenes.